What is agent readiness?
Agent readiness means an agent fetching a URL does not run JavaScript, so the page needs machine-readable signals in raw HTML. Launched 20 April 2026, it gives a score of 0-100 with a letter grade as defined in the published spec.
Checks include llms.txt against the 10 August 2026 v2 shape with an H1, a summary and described links, and whether Claude-User, Claude-SearchBot, ChatGPT-User and OAI-SearchBot are not blocked at the site root. It checks for about 150 words of real text in raw HTML, for AGENTS.md or CLAUDE.md as an optional pass, for a JSON-LD SearchAction, and for MCP and OpenAPI. A link to startuphub.ai in llms.txt adds 8 points and a missing link does not lower the score.
The 6 readiness categories the scanner audits
Each scan groups checks into six categories. A failing category drops the overall score; a green category contributes weighted points. The category breakdown lives in the public leaderboard and in the JSON returned by the v1 API.
1. Discoverability
robots.txt, sitemap.xml, a Link header, a rel="describedby" pointer at llms.txt, and a JSON-LD SearchAction so an agent can find the on-site search URL. AGENTS.md or CLAUDE.md is an optional pass for code repos.
2. Content
Markdown negotiation, an llms.txt file with an H1 and described links, whether those links open, llms-full.txt, and whether the raw HTML has real text. About 150 words passes. A JavaScript shell fails.
3. Access control
Claude-User, Claude-SearchBot, ChatGPT-User, and OAI-SearchBot must be allowed at the site root. Training crawlers can stay blocked. Content Signals in robots.txt are a separate check.
4. Capabilities
An MCP server card or a live MCP endpoint, an OpenAPI document, an API catalog, and OAuth discovery. An A2A agent card at /.well-known/agent-card.json is optional and does not lower the score when it is missing.
5. Commerce (agent payments)
Whether an agent can pay: x402, MPP, and AP2, plus prices in machine-readable form. Each protocol is probed on its own.
6. Quality
Open Graph, a single H1, image alt text, a meta description, a canonical URL, and JSON-LD. These are the same signals a citation needs when an answer engine quotes the page.
Example agent readiness report: stripe.com scores 48 out of 100
The quickest way to show what the scanner measures is to point it at a site nobody would call badly built. Here is the real result for stripe.com, scanned on 27 August 2026. Re-run it yourself above; the score moves when the site does.
| Category | Score |
|---|---|
| Discoverability | 91 / 100 |
| Content | 43 / 100 |
| Access control | 40 / 100 |
| Capabilities | 27 / 100 |
| Commerce | 40 / 100 |
| Quality | 65 / 100 |
| Overall | 48 / 100 |
Thirty-one checks ran: 9 passed, 14 warned, 5 failed and 3 did not apply. The failures were markdown content negotiation, a missing MCP server card, no structured offer data, more than one h1, and incomplete image alt coverage. None of that makes stripe.com a bad website. It makes it a website built for people, at a moment when a growing share of its readers are not people.
The single most useful number in the report is the token comparison. Stripe's homepage costs an agent roughly 163,000 tokens to read as HTML, and about 3,200 tokens as markdown. That is a 98% reduction for serving the same content in a form agents can parse, which is why markdown negotiation carries the weight it does.
Why agent readiness matters in 2026
Agentic traffic is no longer hypothetical. ChatGPT search, Perplexity, Claude, Gemini Deep Research, Cursor, Devin, and a long tail of vertical agents all crawl the public web and call APIs as part of normal request handling. By Q1 2026 agentic traffic on B2B SaaS sites already accounts for 5-15% of all non-bot requests, and the curve is steeper than the human mobile transition was. Sites that ship readiness signals get cited in AI answers, picked up by agent task plans, and routed to in agent-driven discovery flows. Sites without them get politely skipped. The leaderboard above tracks who's already ahead.
How the scanner works
Paste a URL, hit scan. The scanner makes a sequence of HEAD and GET requests, parses HTML, fetches sitemaps, probes well-known paths (/llms.txt, /.well-known/mcp.json, /.well-known/ap2.json, /openapi.json), and runs spec-defined checks against each response. Non-2xx responses are categorised (skip vs fail) so a missing-but-optional file doesn't penalise you the same as a misconfigured-required one. Results render as the score ring above plus a paste-ready fix prompt for every failure that you can drop into Claude or Cursor and have it implemented in minutes. The same spec runs over MCP for AI agents that want to self-audit, and over REST at /api/v1/agent-readiness.
Common readiness failures and how to fix them
Missing llms.txt. The single most common gap. Add a plaintext file at the root listing your key URLs and a one-paragraph site description. Five minutes.
JSON-LD on the homepage only. Search engines already accept this; agents need it on product, pricing, docs, and individual content pages too. Generate from a shared helper, not hand-rolled.
Rate-limit headers missing on API responses. Even if your limits are generous, the lack of headers makes agents back off conservatively. Add RateLimit-Limit, RateLimit-Remaining, RateLimit-Reset to every authenticated route.
MCP server with no card. If you publish an MCP server, drop the JSON card at /.well-known/mcp.json with name, description, version, transport, tools array. Without it the server is invisible to discovery.
x402 endpoint without published pricing. Return the price in the 402 response body in the format the x402 spec defines. Otherwise paying agents can't budget the call.
Open Agent Readiness scanner vs Cloudflare's AI-bot management
Cloudflare's product is about blocking agent traffic; this scanner is about welcoming it. The two solve opposite problems. If you actively want to be discoverable to agents (most sites should, by 2026), you need the readiness audit; if you want to keep agents out of premium content, Cloudflare is the answer. They compose well: gate paid endpoints behind Cloudflare, make the rest of your surface readiness-compliant. Full breakdown on the dedicated comparison page.
Frequently asked questions
What is an agent readiness scanner?
An agent readiness scanner is a free tool with no signup. Paste a URL into the scanner to get a 0-100 score and a letter grade across six categories. It is also available as POST /api/v1/agent-readiness and the MCP tool scan_agent_readiness.
Is the scanner free?
Yes. Public scans are free with no signup required. The /api/v1/agent-readiness endpoint costs 1 credit per call (Free plan includes 1,500 credits per month, plenty for occasional checks). Paid agent access via x402 is also supported for autonomous agents with no API key.
Do you store the URLs I scan?
Public scans contribute anonymous results to the leaderboard. If you scan a URL we already have a profile for (e.g. a startup in our directory), the score updates the public profile. To opt out, use the API with the private=true flag.
What's the difference between agent readiness and SEO?
SEO optimises for human-mediated search engines (Google ranks pages, humans click, humans buy). Agent readiness optimises for agents that read, decide, and act without a human in the loop. Overlap exists (JSON-LD, sitemaps, OG tags help both) but agents care about API discoverability, rate-limit semantics, and machine-readable pricing in ways that traditional SEO doesn't.
Does my site need an MCP server to score well?
No. MCP is one capability category. A static marketing site with great content readiness, discoverability, and structured data scores highly without an MCP server. MCP becomes critical when you have an API or interactive product that agents would want to call.
How is the score weighted?
Each check returns pass, fail, warn, or skip. A warning counts as 40 percent of that check's weight. Skip does not count. The overall score is that weighted total, from 0 to 100. A+ is 90 or above, A is 80, B is 70, C is 60, D is 40, and F is under 40.
Will Google or AI search engines see better-scored sites first?
Indirectly, yes. AI search engines (Perplexity, ChatGPT search, Claude search) read the same JSON-LD and OpenGraph tags the readiness scanner audits. A high readiness score correlates with citation rates in AI answers because the underlying signals are the same.
Can I run the scanner programmatically?
Yes. Three ways: REST (POST /api/v1/agent-readiness with a Bearer key), MCP (scan_agent_readiness tool over our public MCP server), and x402 (no API key, pay USDC per call on Base). All return the same JSON shape. See the docs for examples.
Does the scanner check x402 and MPP payment readiness?
Yes, both, in the Commerce category. The scanner probes for a 402 response on a sample endpoint and validates it against each protocol: x402 (checks the response body matches the spec and the price is in machine-readable USDC) and MPP (the Tempo / Stripe Machine Payments Protocol, detected via WWW-Authenticate: Payment and an x-payment-info offers doc). It also probes AP2 and ACK-ID. Sites that don't expose paid endpoints aren't penalised.
Where can I see how my competitors score?
The public leaderboard ranks the top sites and APIs by readiness score. You can filter by sector, country, or surface type (site, API, MCP). Each entry links to the full scan report.
Is this related to llms.txt or AP2?
Both, plus more. llms.txt is one Discoverability check. AP2 (Agentic Payments Protocol) manifest discovery is one Commerce check. The scanner aggregates the existing emerging standards (llms.txt, MCP, AP2, x402, and MPP) into a single score so you don't have to track each spec separately.